Demystifying Coveo Machine...
Transcript of Demystifying Coveo Machine...
Automated Relevance, Query Suggestions and Recommendations
Richard Tessier
SVP Products
1 © 2017 Coveo Solutions Inc - Proprietary and Confidential
Demystifying Coveo Machine Learning
Cop
yrig
ht ©
20
17 C
oveo
Sol
utio
ns In
c.
All
right
s re
serv
ed. P
ropr
ieta
ry a
nd C
ON
FID
ENTI
AL
1. Intelligent Query Suggest
2. Automatic Relevance Tuning
3. Recommendations
4. Intelligent Term Detection
5. Personalization
Key Use Cases
Optimizesfor future users
Analyzesuser behavior
Learnsbased on success
3 © 2017 Coveo Solutions Inc - Proprietary and Confidential
General Concepts
4
Coveo Machine Learning Flow
Interface & APIs
Transmit interaction events and leverages the models built using these events
Predictive analytics
Wisdom from the interactions’ sumtotalcues likely intent
Machine learning
Data science drives “what you need next” recommendations
Users RelevanceModels-> =->Events
Captured in our Cloud infrastructure and available for ML models processing
Events
5 © 2017 Coveo Solutions Inc - Proprietary and Confidential
7
Machine Learning configuration
8 © 2017 Coveo Solutions Inc - Proprietary and Confidential
Recommendations
9
IntelligentRecommendations
10 © 2017 Coveo Solutions Inc - Proprietary and Confidential
11
Recommendation
Usage Analytics
ML
Previous Users
Record Users ActionsType: Page View and QueryMandatory: Key/Value (identifies one document in one language)Optional: contentType (to recommend a specific type of document)
Build a model(0..n contentTypes)
Query Pipeline
Index
Current User Recommendation RequestActions History (e.g. page view, query)
RecommendationRequest
Query(@Key = @Value)
13 © 2017 Coveo Solutions Inc - Proprietary and Confidential
Query Suggest
14 © 2017 Coveo Solutions Inc - Proprietary and Confidential
15 © 2017 Coveo Solutions Inc - Proprietary and Confidential
16 © 2017 Coveo Solutions Inc - Proprietary and Confidential
Model = list of suggestions candidatesPast queries that were successfulScore based on the relevance indexLanguage-specific models
At query timePartial query compared to all candidates using a distance metric
Candidates are ranked based on their distance metrics and their relevance score
Time constrained: compute for a predetermined duration.
Query suggestion
18© 2017 Coveo Solutions Inc - Proprietary and Confidential
19© 2017 Coveo Solutions Inc - Proprietary and Confidential
20 © 2017 Coveo Solutions Inc - Proprietary and Confidential
Automated Relevance Tuning
21© 2017 Coveo Solutions Inc - Proprietary and Confidential
22 © 2017 Coveo Solutions Inc - Proprietary and Confidential
Favors most popular and pertinent documents for each query
Implicit synonyms detection
Understands common typos
Learns from search sessionsLinks each query to all subsequent clicksLearns the relation strengths between query parts and clicks
Automatic parameter optimization
Automated Relevance Tuning
Query
Clicked SearchResult
… QueryLast Clicked
Search Result
Clicked SearchResult
Clicked SearchResult
Clicked SearchResult
Typical Search Visit
24 © 2017 Coveo Solutions Inc - Proprietary and Confidential
Intelligent Term Detection
Goal is to query the index with long text strings or natural language queriesex: case description
Algorithm extracts the most important terms, using the customer vocabulary frompast queries
Improved relevance over partial match - when enabled automatically boosts documents matching terms
Intelligent Term Detection
Intelligent Term Detection
27 © 2017 Coveo Solutions Inc - Proprietary and Confidential
Questions?
Richard TessierSVP Products
https://www.linkedin.com/in/richardtessier/
28 ©2017CoveoSolutionsInc- ProprietaryandConfidential
Thank you!
Don’t miss tomorrow’s session on Getting the most out of Coveo Machine Learning!